Blog Posts

Car Equipment Management Kills Artificial Intelligence Search Engine Project

Local 3652
Social share icons

Those who have been receiving our weekly news digest know that technical documentation was working on a Proof of Concept (POC) AI Search Engine with MTAIT and Microsoft for a GPT based search for Car Equipment's documentation of over 55,000 documents.

Management cancelled the project due to "uncertainty of significant benefits" even though there was a successful Proof of Concept.

This is unusual because the cancellation appears strategically inconsistent with the original rationale for approval.

1. The Existing Search System Has Clear Operational Deficiencies

A search platform from 2018 that:

  • takes ~12 seconds to load,
  • then another ~12 seconds per page of results,
  • cannot access large portions of historical documentation
  • is quickly approaching its threshold of 65,000 documents even with its Enterprise License
  • is not incremental when indexed

is already imposing measurable productivity costs.

That means management is not comparing:

“working system vs speculative AI”

It is comparing:

“poor-performing legacy retrieval vs materially improved retrieval”

That distinction matters.

In most engineering environments, search latency alone becomes expensive at scale:

If users search dozens of times daily, those delays compound into thousands of lost labor hours annually.

2. The OCR Limitation 

Traditional enterprise search systems depend heavily on:

  • exact text indexing,
  • keyword matching,
  • structured OCR extraction.

Older scanned PDFs often fail because:

  • OCR confidence is poor,
  • diagrams/tables are unreadable by Adobe (but a human can),
  • terminology evolved over decades,
  • document formatting is inconsistent.

An LLM/RAG-based system changes the retrieval paradigm:

  • semantic understanding instead of exact keywords,
  • contextual similarity,
  • tolerance for OCR imperfections,
  • ability to infer intent from incomplete matches.

That is not merely an incremental improvement.
It unlocks previously unusable institutional knowledge.

For organizations with decades of engineering documentation, that can become strategically valuable.

3. The POC Already Demonstrated the Hardest Part

A large percentage of AI projects fail before proving:

  • ingestion feasibility,
  • indexing 
  • no hallucination,
  • integration practicality.

The POC already demonstrated:

  • workable GPT-based retrieval,
  • accessibility of previously inaccessible content,
  • speed in retrieval

Weird huh?